1. Field of the Invention
This invention relates to file server systems, including those file server systems in which it is desired to maintain reliable file system consistency.
2. Related Art
In systems providing file services, such as those including file servers and similar devices, it is generally desirable for the server to provide a file system that is reliable despite the possibility of error. For example, it is desirable to provide a file system that is reliably in a consistent state, regardless of problems that might have occurred with the file server, and regardless of the nature of the file system operations requested by client devices.
One known method of providing reliability in systems that maintain state (including such state as the state of a file system or other set of data structures) is to provide for recording checkpoints at which the system is known to be in a consistent state. Such checkpoints, sometimes called “consistency points,” each provide a state to which the system can retreat in the event that an error occurs. From the most recent consistency point, the system can reattempt each operation to reach a state it was in before the error.
One problem with this known method is that some operations can require substantial amounts of time in comparison with the time between consistency points. For example, in the WAFL file system (as further described in the Incorporated Disclosures), operations on very large files can require copying or modifying very large numbers of file blocks in memory or on disk, and can therefore take a substantial fraction of the time from one consistency point to another. In the WAFL file system, two such operations are deleting very large files and truncating very large files. Accordingly, it might occur that recording a consistency point cannot occur properly while one of these extra-long operations is in progress.
The fundamental requirement of a reliable file system is that the state of the file system recorded on non-volatile storage must reflect only completed file system operations. In the case of a file system like WAFL that issues checkpoints, every file system operation must be complete between two checkpoints. In the earliest versions of the WAFL file system there was no file deletion manager present, thus very large files created a problem as it was possible that such large files could not be deleted between the execution of two consistency checkpoints.
This problem was partially solved in later versions of the WAFL file system, where a file deletion manager was assigned to perform the operation of file deletion, and a consistency point manager was assigned to perform the operation of recording a consistency point. The file deletion manager would attempt to resolve the problem of extra-long file deletions by repeatedly requesting more time from the consistency point manager, thus “putting off” the consistency point manager until a last-possible moment. However, at that last-possible moment, the file deletion manager would be required to give way to the consistency point manager, and allow the consistency point manger to record the consistency point. When this occurred, the file deletion manager would be unable to complete the file deletion operation. In that earlier version of the WAFL file system, instead of completing the file deletion operation, the file deletion manager would move the file to a fixed-length “zombie file” list to complete the file deletion operation. At a later time, a zombie file manager would re-attempt the file deletion operation for those files on the fixed-length zombie file list.
While this earlier method achieved the general result of performing file deletions on very large files, it has the drawbacks that it is a source of unreliability in the file system. First, the number of files that could be processed simultaneously as zombie files was fixed in the previous version.
Second, the file deletion manager and crash recovery mechanism did not communicate. The file deletion manager did not notify the crash recovery mechanism that a file was being turned into a zombie and the crash recovery mechanism was unable to create zombie files. Thus, to allow a checkpoint to be recorded, a long file would have to be turned into a zombie. If the system crashed at this point, the crash recovery mechanism might not be able to correctly recover the file system since it is unaware that a zombie file should be created and was incapable of creating zombie files should the need arise. Similarly, the operations of the file deletion manager when creating zombie files, and its operations in deleting those zombie files, were not recorded in non-volatile storage, and thus could not be “replayed” after recovery to duplicate the operations of the file deletion manager.
Third, since the file deletion manager and replay mechanism did not communicate, the free space reported could be inaccurately reported. Attempts to restore state could fail, because the amount of free space could be different than that actually available. Attempts to restore state could also fail because the operations of the file deletion manager in using zombie files were not recorded in non-volatile storage; as a result, it might occur that other operations performed during replay could conflict with the file deletion manager and cause a crash.
Fourth, the earlier method is non-deterministic in the sense that it is not assured whether any particular file deletion operation will be completed before or after a selected consistency point. Moreover, the earlier method does not resolve problems associated with other extra-long file operations, such as requests to truncate very large files to much smaller length.
Accordingly, it would be advantageous to provide a technique for extra-long operations in a reliable state-full system (such as a file system) that is not subject to the drawbacks of the known art. Preferably, in such a technique, those parts of the system responsible for recording of consistency points are fully aware of the intermediate states of extra-long operations, the performance of extra-long operations is relatively deterministic, and performance of extra-long operations is atomic with regard to consistency points.